Two AI agents are racing through a maze that periodically rewrites its own walls. Every time it does, both of them stop, reassess, and recalculate — and both are quietly getting better at spotting the kinds of dead ends they've fallen for before. Every run generates a different maze size, pace, and pair of agents. Press F anytime to go fullscreen.
Next reshape in --s
Every few seconds, a wall opens somewhere and another one closes. Neither agent gets a warning; they discover it on their next step, the way a room can feel different the moment you walk back into it. Watch what happens right after a reshape — both agents pause briefly before moving again. That pause isn't hesitation. It's them checking what's actually true now, instead of acting on what used to be true a moment ago.
Every trap marker you see on the board is not a failed run, it's a completed observation. The agent that walks a corridor to its end and turns around hasn't wasted the trip — it now knows something about that corridor's shape it didn't know thirty seconds ago. The backtrack count next to each agent's name isn't a scoreboard of mistakes. It's closer to a record of how much ground has actually been checked.
The maze regenerates completely between rounds, so no single memorized route survives for long. What does survive, held quietly in the background, is the shape of the fork that tended to lead nowhere — a certain arrangement of open and closed walls at a junction. That shape can reappear in a maze built from entirely different randomness. Noticing it there is a small but real form of learning: not "I've been here before," but "I've seen something built like this before."
Neither agent avoids every dead end, and neither should — a maze with no wrong turns wouldn't be much of a maze. The more telling number on this board is how quickly each agent gets moving again after the ground shifts, and how often the "traps avoided" count ticks up over a long run. Resilience, here and probably elsewhere, looks less like never hitting a wall and more like getting steadily faster at recognizing which walls you've hit before.